CN106022373B - A kind of image-recognizing method based on extended mean value canonical correlation analysis - Google Patents
A kind of image-recognizing method based on extended mean value canonical correlation analysis Download PDFInfo
- Publication number
- CN106022373B CN106022373B CN201610331173.0A CN201610331173A CN106022373B CN 106022373 B CN106022373 B CN 106022373B CN 201610331173 A CN201610331173 A CN 201610331173A CN 106022373 B CN106022373 B CN 106022373B
- Authority
- CN
- China
- Prior art keywords
- correlation analysis
- mean value
- canonical correlation
- sample
- gmcca
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of robust canonical correlation analysis algorithm based on extended mean value, mainly solves the problems, such as that traditional canonical correlation analysis algorithm based on Euclidean distance causes sample covariance singular problem to outlier non-robust and higher-dimension small sample.Realization process are as follows: (1) input necessary parameter, and centralization processing is carried out to training sample；(2) two groups of set of projections of traditional canonical correlation analysis are solved；(3) based on the objective optimization function of extended mean value reconstruction model, to inhibit influence of the outlier to objective function；(4) objective function is solved with linear iterative method, wherein two groups of set of projections using traditional canonical correlation analysis are initialized；(5) two groups of set of projections for acquiring the canonical correlation analysis based on extended mean value are used for feature extraction and the dimensionality reduction of sample.Experiment show on multiple features handwritten form database (MFD), face database (ORL) and object image data library (COIL-20) validity of the algorithm.
Description
Technical field
The invention belongs to feature extraction and data dimensionality reduction technology field, the improvement of predominantly canonical correlation analysis algorithm is excellent
Change.Specially a kind of image-recognizing method based on extended mean value canonical correlation analysis can be applied to machine learning, mode is known
Not, the fields such as data mining and image procossing.
Background technique
In pattern-recognition and machine learning field, Dimensionality Reduction (Dimensionality reduction, DR) is always
One of hot topic of research.There has been proposed the dimensionality reduction that a large amount of method is used for data, wherein principal component analysis (Principal
Component analysis, PCA) it is one of most classic method.But PCA is concerned with single mode data (Single-view
Data feature extraction and dimensionality reduction).With the development of science and technology, the means that people acquire data are more diversified, same thing
With polynary expression form, such as someone information can be made of attributes such as face, writing, fingerprints.For multi-modal number
According to (Multi-view data), canonical correlation analysis (Canonical correlation analysis, CCA) is more suitable
The extraction of feature with merge.CCA is a kind of multivariate statistical method for studying correlation between two groups of variables of same target, be can be used for
Realize feature extraction, dimensionality reduction and the visualization of data.CCA by maximize different modalities between correlation, eliminate data between
Redundancy extracts important feature, the performance of strengthening subsequent study (as classified) task.In recent years, CCA and its derivative model at
Function is applied to the fields such as recognition of face, meteorologic analysis, biological information fusion and social science.But CCA is substantially a kind of linear
The learning method of subspace, what is learnt is the linear character under a kind of global linear case.For nonlinear scene,
CCA study often leads to owe the result of study.For this purpose, S.Akaho combination nuclear technology propose core CCA (Kernel CCA,
KCCA), deficiency of the CCA under nonlinear situation is overcome.2000, San et al., which is proposed, was locally linear embedding into (Local
Linear embedding, LLE) Method of Nonlinear Dimensionality Reduction, manifold learning (Manifold learning, ML) obtains from this
In-depth study.Sun et al. be introduced into locality preserving projections in manifold learning (Locality preserving projection,
LPP thought) retains the manifold structure information in data, proposes a kind of CCA (Locality preserving locally kept
CCA, LPCCA), application of the CCA under nonlinear situation has been expanded significantly.Nevertheless, CCA, KCCA and LPCCA are all based on
The method of Euclidean distance, from the point of view of multiple linear regression analysis, their objective optimization function is all based on L2 norm
Least mean-square error (Mean square error, MSE).However, outlier is prevalent in observation in reality scene
In data set.Studies have shown that using the Euclidean distance method of MSE of L2 norm, for outlier, all there is non-robust.And
And CCA, KCCA and LPCCA are eventually converted into generalized eigenvalue solution, under higher-dimension Small Sample Size, sample covariance square
Battle array is most probably unusual, this affects the robustness of algorithm.
Extended mean value (Generalized mean, GM) is the popularizing form of arithmetic mean of instantaneous value.P value by adjusting GM can
To show the center of a variety of data.J.Oh et al. 2013 in document (Generalized mean for feature
extraction in one-class classification problems.Pattern Recognition,2013,46
(12): 3328-3340 in conjunction with being proposed based on extended mean value, one kind is novel inclined discriminatory analysis (Biased in)
Discriminant analysis using generalized mean, BDAGM), enhance the effect of positive sample, inhibits wild
It is worth the interference of point.It is demonstrated experimentally that GM enhances the robustness of algorithm, the performance of algorithm is improved face.
Summary of the invention
It is generally existing to outlier non-robust and height in order to solve the canonical correlations algorithm such as traditional CCA, KCCA and LPCCA
Small sample problem is tieed up, is affected when causing to extract characteristics of image by noise and image pattern amount are few, to reduce identification
The problem of rate, the present invention propose a kind of image-recognizing method based on extended mean value canonical correlation analysis, and the method is based on broad sense
The robust canonical correlation analysis algorithm (CCA based on generalized mean, GMCCA) of mean value carries out image recognition.
In GMCCA algorithm, the correlated error concept in projector space between sample is proposed first, to better describe the sample after projection
Between similarity degree；Secondly, being based on extended mean value, objective function is rebuild by correlated error, is replaced original based on L2
The objective function of the least mean-square error of norm, obtains new model；Finally, the method by linear iteraction solves new model.?
Multiple features handwritten form database (Multiple feature database, MFD), human face data collection (ORL) and characteristics of objects number
Showing new algorithm not only according to the experiment in library (COIL-20) three real data sets has better robustness, but also avoids
The problem that higher-dimension small sample causes sample covariance matrix unusual.A kind of image recognition based on extended mean value canonical correlation analysis
The step of method, specifically can be described as follows:
(1) image pattern is collected；
(2) sample set that one group of size is N is inputtedThe parameter p of extended mean value, it is interior
Portion iteration total degree T1And T2, outer iteration total degree T, the dimension d of feature after dimensionality reduction；
(3) firstly, calculating sample set X=(x1,x2,...,xN) and Y=(y1,y2,...,yN) central value:
It is used in combinationWithCentralization X and Y:
For uniformity, after centralizationWithStill it is denoted as X=(x1,x2,...,xN) and Y=(y1,y2,...,
yN)；
(4) traditional canonical correlation analysis (Canonical correlation analysis, CCA) is to find two groups of samples
The projection vector of this collection X and YWithSo that in projector space two groups of sample sets feature have maximum
Correlation, criterion function is as follows:
Above formula is converted into two following eigen characteristic value problems:
And wxAnd wyWith following equilibrium relationships:
Sxywy=λ Sxxwx,Syxwx=λ Syywy
Combination of eigenvectors corresponding to maximum preceding d eigen characteristic value is finally chosen into two groups of set of projectionsWith
As can be seen that the solution of CCA is needed to SxxAnd SyyIt inverts.But higher-dimension small sample easily leads to SxxAnd SyyIt is unusual, shadow
Ring the performance of CCA；
(5) p ≠ 0 is assumed, for a scalar datasets { ai> 0, i=1,2 ..., N extended mean value MGBe defined as
Under:
Further analysis, extended mean value MGInIt can be by data set { aiOne group of non-negative linear combination table
Show, as follows:
biA can be regarded asiWeight, i.e. aiTo MGContribution margin.As p < 1, with aiIt is bigger, biIt is smaller, it is meant that when
When p < 1, extended mean value MGBy { aiIn smaller value be affected, and p is smaller, influences bigger.This property of extended mean value
Matter plays main function in the influence that GMCCA inhibits outlier.
Correlated error e (W in projector space shown in being defined as follows between samplex,Wy):
In conjunction with above-mentioned extended mean value and correlated error, the robust canonical correlation analysis based on extended mean value as follows is constructed
The objective optimization function of (CCA based on generalized mean, GMCCA):
Above-mentioned objective function is solved, obtains WxAnd Wy.Take WxAnd WyTwo groups of set of projections of d column composition GMCCA before obtainingWithThe essence of GMCCA robustness as can be seen from the above equation: as p < 1, αiValue with relative error increase and
Reduce, therefore, sample point biggish for correlated error in projector space, i.e. outlier impart lesser weight, inhibit wild
Adverse effect of the value point to criterion function, enhances the robustness of algorithm；
(6) it is acquired using step (5)WithFeature extraction and dimensionality reduction are carried out to original sample:
It willWithFor next pattern recognition task；
(7) identification mission of image is completed using nearest neighbor classifier.
Above-mentioned objective function is solved by a kind of linear iterative method, and this method is specific as follows:
Assuming that the number t of current iteration1=t2=t=0；The W that the t times iteration obtainsxAnd WyRespectivelyWithAnd
InitializationWith
It is fixed firstIt is acquired by following minimum problem:
I.e.It isThe corresponding orthogonal eigenvectors collection of maximum d characteristic value；
At this point, withIt goes to update Wx.It is fixedSimilarly,It is acquired by following minimum problem:
So far it can obtain, solve WxAnd WyLinear iterative algorithm it is as follows:
As can be seen that GMCCA is different from traditional CCA from above-mentioned linear iterative method, two groups of features of GMCCA are thrown
Photograph album is to separate to solve to obtain, WxAnd WyHave no the equilibrium relationships in CCA.WxAnd WyIt is the weighting association of sample set X and Y respectively
The corresponding orthogonal eigenvectors collection of the maximum d characteristic value of variance.Moreover, entire solution procedure is not related to sample set X and Y
Covariance matrix invert.Therefore, GMCCA avoids the higher-dimension small sample in traditional CCA and causes sample covariance matrix odd
Different problem
GMCCA algorithm used in the present invention has the advantage that
(1) influence of the outlier to objective optimization function is inhibited by extended mean value.
(2) the sample rotational invariance of Euclidean distance is remained.
(3) GMCCA avoids the problem that high dimensional and small sample size problem causes sample covariance matrix unusual.
And a kind of image-recognizing method based on extended mean value canonical correlation analysis provided by the invention is mentioned in characteristics of image
Strong robustness when taking can preferably cope with picture noise and the few problem of image pattern amount, so having higher discrimination.
Detailed description of the invention
Fig. 1 is GMCCA algorithm implementation flow chart in the present invention；
Fig. 2 is the 6 width images of a people in ORL face database；
Fig. 3 is 4 width images training before choosing the every class of ORL, and GMCCA and other 4 kinds of algorithms are in the case where O-L feature is combined with dimension
The recognition result of variation；
Fig. 4 is 4 width images training before choosing the every class of ORL, and GMCCA and other 4 kinds of algorithms are in the case where O-H feature is combined with dimension
The recognition result of variation；
Fig. 5 is 4 width images training before choosing the every class of ORL, and GMCCA and other 4 kinds of algorithms are in the case where L-H feature is combined with dimension
The recognition result of variation；
Fig. 6 is 20 object images in COIL-20；
Fig. 7 is 25 width images training before choosing the every class of COIL-20, and GMCCA is with other 4 kinds of algorithms in the case where O-L feature combines
The recognition result changed with dimension；
Fig. 8 is 25 width images training before choosing the every class of COIL-20, and GMCCA is with other 4 kinds of algorithms in the case where O-H feature combines
The recognition result changed with dimension；
Fig. 9 is 25 width images training before choosing the every class of COIL-20, and GMCCA is with other 4 kinds of algorithms in the case where L-H feature combines
The recognition result changed with dimension；
Specific embodiment
In order to illustrate the object, technical solutions and advantages of the present invention, below in conjunction with specific embodiments and drawings, to the present invention
It is described in further details.
Referring to Fig.1, specific implementation process of the invention the following steps are included:
(1) image pattern is collected；
(2) sample set that one group of size is N is inputtedThe parameter p of extended mean value, it is interior
Portion iteration total degree T1And T2, outer iteration total degree T, the dimension d of feature after dimensionality reduction；
(3) firstly, centralization sample set X=(x1,x2,...,xN) and Y=(y1,y2,...,yN):
Calculate sample set X=(x1,x2,...,xN) and Y=(y1,y2,...,yN) central value:
It is used in combinationWithCentralization X and Y:
For uniformity, after centralizationWithStill it is denoted as X=(x1,x2,...,xN) and Y=(y1,y2,...,yN)；
(4) objective function of tradition CCA is as follows:
Above formula is converted into two following eigen characteristic value problems:
Combination of eigenvectors corresponding to maximum preceding d eigen characteristic value is finally chosen into two groups of set of projectionsWith
(5) p ≠ 0 is assumed, for a scalar datasets { ai> 0, i=1,2 ..., N extended mean value MGBe defined as
Under:
And the correlated error e (W in projector space shown in being defined as follows between samplex,Wy):
In conjunction with above-mentioned extended mean value and correlated error, the robust canonical correlation analysis based on extended mean value as follows is constructed
The objective optimization function of (CCA based on generalized mean, GMCCA):
Above-mentioned objective optimization function is solved by a kind of linear iterative method, and this method is specific as follows:
Assuming that the number t of current iteration1=t2=t=0；The W that the t times iteration obtainsxAnd WyRespectivelyWithAnd
InitializationWith
It is fixed firstIt is acquired by following minimum problem:
I.e.It isThe corresponding orthogonal eigenvectors collection of maximum d characteristic value；
At this point, withIt goes to update Wx.It is fixedSimilarly,It is acquired by following minimum problem:
So far it can obtain, solve WxAnd WyLinear iterative algorithm it is as follows:
Take WxAnd WyTwo groups of set of projections of d column composition GMCCA before obtainingWith
(6) it is acquired using step (5)WithFeature extraction and dimensionality reduction are carried out to original sample:
It willWithFor next pattern recognition task；
(7) identification mission of image is completed using nearest neighbor classifier.
Effect of the invention can be further illustrated by the experiment below on truthful data library.
1. description of test
For the validity for verifying GMCCA, this section is in multiple features handwritten form database (Multiple feature
Database, MFD), it carries out in three real data sets of human face data collection (ORL) and characteristics of objects database (COIL-20) real
It tests, and induces CCA with PCA, CCA, robust CCA (Robust CCA, ROCCA), complete CCA (Complete CCA, C3A) and core
(CCA based on kernel-induced measure, KI-CCA) is compared.ROCCA passes through building approximate matrix generation
For sample covariance matrix, eliminate high dimensional and small sample size problem, the experimental verification of the identification validity of ROCCA.C3A overcomes
The problem of CCA may lose information, extracts more complete canonical correlation information.ROCCA induces distance metric generation with core
For the euclidean distance metric of traditional CCA, while improving algorithm robustness, and nonlinear problem is solved.
In all experiments herein, the p of GMCCA is set as 0.1, t1、t210,10 and 20 are respectively set to T.PCA needs
2 groups of features are joined end to end to form new high dimensional feature vector, then carry out feature extraction with PCA, CCA, ROCCA, C3A,
By concatenated mode after KICCA and GMCCA extraction feature, i.e., the feature after two groups of dimensionality reductions is serially connected end to end
Carry out discriminance analysis.Classifier uses nearest neighbor classifier.
2. experimental result
Test the experiment of 1 multiple features handwritten form
The performance of GMCCA is tested in this experiment with the hand-written volumetric data set of multiple features (MFD).The data set is UCI machine learning
One component part (http://archive.ics.uci.edu/ml/datasets/Multiple+ of knowledge base
Features), there is important value in Handwritten Digital Recognition.The database includes 0~9 totally 10 digital 6 features
Data set, 200 samples of every class, totally 2000 samples, are widely used in the research of pattern-recognition and machine learning.From two-value
Change and extract 6 features in handwriting digital image, including Fourier coefficient, profile correlated characteristic, Karhunen-Loeve expansion
Feature, pixel be average, Zernike square and morphological feature, corresponding feature name and dimension are respectively as follows: (fou, 76),
(fac, 216), (kar, 64), (pix, 240), (zer, 47) and (mor, 6).On this data set, optional 2 groups of feature conducts
Input, shares 15 kinds of combinations.Each feature is combined, 100 samples are randomly selected from every class as training, are left
100 samples as test.
Table 1 show the average recognition result for 10 random experiments that 6 kinds of algorithms are closed in different characteristic group, every kind of algorithm
In best identified rate indicated with black matrix, similarly hereinafter.It can be seen that in most of combination from result shown in table
The average recognition rate of GMCCA algorithm is better than other algorithms, is especially apparent the recognition effect higher than CCA, in addition, 15 kinds of combinations is flat
Equal discrimination is also above other algorithms.These result verifications validity of GMCCA.In fou-pix, kar-pix, mor-pix
In mor-zer combination, the discrimination of GMCCA is also illustrated that lower than other algorithms although the discrimination of GMCCA is still higher than CCA
GMCCA still has shortcoming in the combination of some features.
16 kinds of algorithms of table recognition result that different characteristic group is closed in MFD experiment
Test the experiment of 2 ORL face databases
In order to further verify the validity of GMCCA, the ORL database that human face posture changes greatly is chosen in this experiment
(http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html).The database
It is made of the laboratory Britain Camb Olivetti from April, 1992 to a series of facial images shot during in April, 1994,
Share 40 all ages and classes, different sexes and not agnate object.Each 10 width image of object amounts to 400 width gray level image groups
At picture size is 92 × 112, and image background is black.Wherein face part expression and details change, for example, laugh at not
Laugh at, eyes are opened or closed, and are worn or are not worn glasses, human face posture also changes, depth rotation and Plane Rotation up to 20
Degree, facial size also have most 10% variation.The library is current most popular standard database, it contains a large amount of ratio
Relatively result.Fig. 2 shows the 6 width images of a people in ORL database.
4,5,6,7 or 8 width images are randomly selected in experiment from everyone 10 width images as training, remaining, which is used as, surveys
Examination；3 groups of features are extracted to each image.Wherein, primitive image features are denoted as O；By original image local binary patterns
Feature after (Local binary pattern, LBP) is extracted is denoted as L；By original image histograms of oriented gradients
Feature after (Histogram of Oriented Gradient, HOG) is extracted is denoted as H.LBP and HOG feature and combinations thereof is special
Sign has proved to be effective in recognition of face problem.In order to avoid singularity problem, with PCA by above-mentioned 3 kinds of feature reducings
To 100 dimensions.
Table 2 show the average recognition result for 10 random experiments that 6 kinds of algorithms are closed in 3 kinds of feature groups, and " n " indicates every
The number of training of class, similarly hereinafter.From table 2 it can be seen that the recognition effect of GMCCA is in 3 kinds of various combinations in the case where number big absolutely
It is superior to other 5 kinds of algorithms, and more all combined average recognition rates, GMCCA is also superior to other 5 kinds of algorithms.From table 2
It can also be seen that the recognition effect of GMCCA improves a lot than traditional CCA, and especially when number of training is less, such as every class
4 training samples.These are the result shows that the feature of GMCCA extraction demonstrates the validity of method with more robustness.In table 2
Result also show, there are four types of in the case of, the discrimination of GMCCA is slightly below other algorithms, but very close with optimal value.
Recognition result of the 26 kinds of algorithms of table on ORL face database
The preceding 4 width image for choosing everyone in ORL database again is trained, and residual image is for testing, experimental result
As shown in figure 3, figure 4 and figure 5.Because other 5 kinds of algorithms are substantially better than PCA in table 2, Fig. 2, Fig. 3 and Fig. 4 illustrate only 5
Kind canonical correlation algorithm is in the case where 3 kinds of features are combined with the recognition result of dimension variation.It can be seen that from Fig. 2, Fig. 3 and Fig. 4 that GMCCA
Better than other 4 kinds of algorithms, especially in the case where dimension is less, the discrimination of GMCCA is apparently higher than other algorithms.From algorithm
Stability angle, GMCCA are also preferable than other 4 kinds of algorithms.Experimental result effectively demonstrates the robustness of GMCCA again.
Test the experiment of 3 COIL-20 object databases
Using the COIL-20 object database being widely used in the world, COIL-20 is Columbia University for this section experiment
One include 20 objects image data base (http://www.cs.columbia.edu/CAVE/software/
Softlib/coil-20.php), the database respectively to each object from 0 °~360 ° carry out horizontal direction rotation, every 5 °
Piece image is sampled, each object is total to take 72 width images, amounts to 1440 width images.The database has been successfully applied for
The fields such as pattern-recognition and machine learning, such as the visualizations of data, the estimation of posture.20 in COIL-20 database are right
As shown in Figure 6.
In experiment, 10,20,30,40 and 50 width images, remaining image are randomly selected from 72 width images of each object
As test.It is independent to carry out 10 random experiments, then calculate its average recognition rate.3 groups of spies are extracted to each image in experiment
Sign.This experiment will primitive image features be denoted as O；Feature after original image is extracted with LBP is denoted as L；Original image is used
Feature after HOG is extracted is denoted as H.And PCA is executed by above-mentioned 3 kinds of feature reducings to 50 dimensions.
Table 3 shows the average recognition result for 10 random experiments that 6 kinds of algorithms are closed in 3 kinds of feature groups.From the reality of table 3
It tests result and can be seen that GMCCA and be substantially better than traditional CCA.In most cases, GMCCA is slightly better than ROCCA effect.
In table 3, the discrimination of CCA and C3A are suitable, illustrate this data set after PCA extracts Feature Dimension Reduction, CCA can extract complete
Characteristic information, and the discrimination of GMCCA be better than CCA and C3A, also illustrate that GMCCA not only extracts complete characteristic information, and
And the feature extracted more has robustness.In table 3 still there are two types of, the discrimination of GMCCA is than the summary of other algorithms
It is low, but difference very little.Also, from the point of view of ensemble average discrimination, GMCCA is better than other 5 kinds of algorithms.These experiment shows
GMCCA validity and robustness.
Recognition result of the 36 kinds of algorithms of table on COIL-20 object database
The preceding 25 width image for choosing each object in COIL-20 database again is trained, and residual image is used to test,
Fig. 7, Fig. 8 and Fig. 9 show algorithm in 5 in the case where 3 kinds of features are combined with the recognition result of dimension variation.It can from the result of 3 figures
To find out that GMCCA is substantially better than other 4 kinds of algorithms, compared to traditional CCA, discrimination is enhanced, and further
Demonstrate GMCCA discrimination when dimension is less conclusion more higher than other algorithms.Moreover, increase of the GMCCA with dimension, identification
Rate more tends towards stability than other 4 kinds of algorithms, and the feature that these results illustrate that GMCCA is extracted more has robustness.It is noted that
The Dimension-Recognition Rate broken line of CCA and C3A is to be overlapped, and demonstrates the discrimination of CCA and C3A in table 3
Comparable conclusion illustrates that CCA can extract complete characteristic information from data set.This is also reflected in CCA and can mention in side
While taking standby information, GMCCA can inhibit the influence of outlier, extract more robust feature.Above-mentioned experimental result
Further demonstrate the validity and robustness of GMCCA.
Claims (2)
1. a kind of image-recognizing method based on extended mean value canonical correlation analysis, comprising the following steps:
(1) image pattern is collected；
(2) sample set that one group of size is N is inputtedxi,yiIndicate two width of i-th of sample
Image；The parameter p of extended mean value, inner iterative total degree T1And T2, outer iteration total degree T, the dimension d of feature after dimensionality reduction；
(3) firstly, calculating sample set X=(x1,x2,...,xN) and Y=(y1,y2,...,yN) central value:
It is used in combinationWithCentralization X and Y:
For uniformity, after centralizationWithStill it is denoted as X=(x1,x2,...,xN) and Y=(y1,y2,...,yN)；
(4) two groups of set of projections of traditional canonical correlation analysis (Canonical correlation analysis, CCA) are obtainedWithCCA is the projection vector for finding two groups of sample sets X and YWithSo that in projector space
The features of two groups of sample sets there is maximum correlation, criterion function is as follows:
Above formula (0) is converted into two following eigen characteristic value problems:
And wxAnd wyWith following equilibrium relationships:
Sxywy=λ Sxxwx,Syxwx=λ Syywy
Combination of eigenvectors corresponding to maximum preceding d eigen characteristic value is finally chosen into two groups of set of projectionsWith
(5) p ≠ 0 is assumed, for a scalar datasets { ai> 0, i=1,2 ..., N extended mean value MGIt is defined as:
And the correlated error e (W in projector space shown in being defined as follows between samplex,Wy):
In conjunction with above-mentioned extended mean value and correlated error, the robust canonical correlation analysis (CCA based on extended mean value as follows is constructed
Based on generalized mean, GMCCA) objective optimization function:
Above-mentioned objective optimization function is solved, obtains WxAnd Wy；Take WxAnd WyTwo groups of set of projections of d column composition GMCCA before obtainingWith
(6) it is acquired using step (5)WithFeature extraction and dimensionality reduction are carried out to original sample:
It willWithFor next pattern recognition task；
(7) identification mission of image is completed using nearest neighbor classifier.
2. a kind of image-recognizing method based on extended mean value canonical correlation analysis as described in claim 1, in step (5)
The objective optimization function of GMCCA is by following conversion:
Wherein, | | it is ABS function, guarantees αiNonnegativity；Above formula is solved using the method for linear iteraction, by following
Step carries out:
The number t of (5-1) hypothesis current iteration1=t2=t=0；The W that the t times iteration obtainsxAnd WyRespectivelyWithAnd
InitializationWith
(5-2) is fixed first It is acquired by following minimum problem:
In above formula (1)It is the weighting covariance matrix of sample set Y,It isMaximum d characteristic value is corresponding orthogonal
Set of eigenvectors；
At this point, withIt goes to update Wx；It is fixedSimilarly,It is acquired by following minimum problem:
So far it can obtain, solve WxAnd WyLinear iterative algorithm it is as follows:
(5-3) is according to step (5-2) as a result, two groups of set of projections of GMCCA are
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610331173.0A CN106022373B (en) | 2016-05-18 | 2016-05-18 | A kind of image-recognizing method based on extended mean value canonical correlation analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610331173.0A CN106022373B (en) | 2016-05-18 | 2016-05-18 | A kind of image-recognizing method based on extended mean value canonical correlation analysis |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106022373A CN106022373A (en) | 2016-10-12 |
| CN106022373B true CN106022373B (en) | 2019-04-23 |
Family
ID=57098943
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610331173.0A Active CN106022373B (en) | 2016-05-18 | 2016-05-18 | A kind of image-recognizing method based on extended mean value canonical correlation analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106022373B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106845358B (en) * | 2016-12-26 | 2020-11-10 | 苏州大学 | Method and system for recognizing image features of handwritten characters |
| CN108876715B (en) * | 2018-05-24 | 2021-06-01 | 海南大学 | Image data robust bilateral 2D linear discriminant analysis dimension reduction method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20080100097A (en) * | 2007-05-11 | 2008-11-14 | 포항공과대학교 산학협력단 | Method for recognizing face using two-dimensional canonical correlation analysis |
| JP2009133867A (en) * | 2001-08-13 | 2009-06-18 | Bg Medicine Inc | Method and system for profiling biological system |
| CN102289679A (en) * | 2011-08-30 | 2011-12-21 | 西安交通大学 | Method for identifying super-resolution of face in fixed visual angle based on related characteristics and nonlinear mapping |
| CN103218606A (en) * | 2013-04-10 | 2013-07-24 | 哈尔滨工程大学 | Multi-pose face recognition method based on face mean and variance energy images |
| CN103745205A (en) * | 2014-01-22 | 2014-04-23 | 山东大学 | Gait recognition method based on multi-linear mean component analysis |
-
2016
- 2016-05-18 CN CN201610331173.0A patent/CN106022373B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009133867A (en) * | 2001-08-13 | 2009-06-18 | Bg Medicine Inc | Method and system for profiling biological system |
| KR20080100097A (en) * | 2007-05-11 | 2008-11-14 | 포항공과대학교 산학협력단 | Method for recognizing face using two-dimensional canonical correlation analysis |
| CN102289679A (en) * | 2011-08-30 | 2011-12-21 | 西安交通大学 | Method for identifying super-resolution of face in fixed visual angle based on related characteristics and nonlinear mapping |
| CN103218606A (en) * | 2013-04-10 | 2013-07-24 | 哈尔滨工程大学 | Multi-pose face recognition method based on face mean and variance energy images |
| CN103745205A (en) * | 2014-01-22 | 2014-04-23 | 山东大学 | Gait recognition method based on multi-linear mean component analysis |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106022373A (en) | 2016-10-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tuia et al. | Semisupervised manifold alignment of multimodal remote sensing images | |
| Wu et al. | A comprehensive study on cross-view gait based human identification with deep cnns | |
| Yuan et al. | Hyperspectral image classification via multitask joint sparse representation and stepwise MRF optimization | |
| Chen et al. | An asymmetric distance model for cross-view feature mapping in person reidentification | |
| Huerta et al. | A deep analysis on age estimation | |
| Zhou et al. | Hierarchical gaussianization for image classification | |
| Patel et al. | Dictionary-based face recognition under variable lighting and pose | |
| Yang et al. | Kernel ICA: An alternative formulation and its application to face recognition | |
| Liao et al. | An improvement to the SIFT descriptor for image representation and matching | |
| CN104392246B (en) | It is a kind of based between class in class changes in faces dictionary single sample face recognition method | |
| Yu et al. | Can we see more? Joint frontalization and hallucination of unaligned tiny faces | |
| CN108647690A (en) | The sparse holding projecting method of differentiation for unconstrained recognition of face | |
| Jiang et al. | Patch‐based principal component analysis for face recognition | |
| Rodolà et al. | SHREC'17: Deformable shape retrieval with missing parts | |
| Hidru et al. | EquiNMF: Graph regularized multiview nonnegative matrix factorization | |
| CN106096517A (en) | A kind of face identification method based on low-rank matrix Yu eigenface | |
| CN111414958B (en) | Multi-feature image classification method and system for visual word bag pyramid | |
| Li et al. | Online metric-weighted linear representations for robust visual tracking | |
| CN109241813A (en) | The sparse holding embedding grammar of differentiation for unconstrained recognition of face | |
| CN111325275A (en) | Robust image classification method and device based on low-rank two-dimensional local discriminant map embedding | |
| CN111723759B (en) | Unconstrained face recognition method based on weighted tensor sparse graph mapping | |
| Luevano et al. | A study on the performance of unconstrained very low resolution face recognition: Analyzing current trends and new research directions | |
| Chandaliya et al. | Child face age progression and regression using self-attention multi-scale patch gan | |
| CN106022373B (en) | A kind of image-recognizing method based on extended mean value canonical correlation analysis | |
| Zhang et al. | Face Recognition Using L-Fisherfaces. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20210922 Address after: 710016 B 1616, Tiandi Times Plaza, Fengcheng two road, Weiyang District, Xi'an, Shaanxi. Patentee after: Liu Jiaojiao Address before: No. 1800 road 214122 Jiangsu Lihu Binhu District City of Wuxi Province Patentee before: Jiangnan University |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20211115 Address after: 710000 rooms 513 and 514, block C, Boyuan science and Technology Plaza, Xi'an Jiaotong University Science and Technology Park, Yanta District, Xi'an City, Shaanxi Province Patentee after: Hangao Aviation Technology (Xi'an) Co., Ltd Address before: 710016 1616, block B, Tiandi Times Square, Fengcheng Second Road, Weiyang District, Xi'an City, Shaanxi Province Patentee before: Liu Jiaojiao |